The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes data from external sources such as GitHub profiles, ArXiv papers, and web search results to generate email outreach and populate trackers.\n
Ingestion points: Untrusted data enters the agent context via web_search_live, person_search, and web_enrich_live tool calls described in SKILL.md.\n
Boundary markers: The instructions lack explicit delimiters or warnings to ignore embedded instructions within the ingested candidate data (e.g., in bios or headlines) before interpolation into emails.\n
Capability inventory: The agent utilizes capabilities including script execution via the execute tool and email draft creation via gmail_create_draft.\n
Sanitization: While the CSV logging uses csv.DictWriter, the email body generation lacks specific sanitization to prevent the agent from reacting to malicious instructions hidden in the candidate's technical output or profile metadata.\n- [DYNAMIC_EXECUTION]: The skill uses dynamic code generation to interact with APIs and manage local files.\n
Evidence: SKILL.md provides multiple JavaScript templates for the agent to generate and run via the execute tool to perform complex database queries and enrichment.\n
Evidence: Phase 5 in SKILL.md generates a Python script to handle CSV logging for the outreach tracker.

candidate-sourcing